CerviDBF-Net: Leveraging Spatial and Temporal Features for Accurate Cervical Cancer Diagnosis

Abstract

Cervical cancer screening increasingly relies on automated image analysis to support colposcopic assessment and improve the consistency of clinical decision-making. This study proposes CerviDBF-Net, a novel deep learning framework for automated cervix-type classification from colposcopic images. The proposed architecture combines DenseNet121 for spatial feature extraction, a brain-inspired Liquid Neural Network (LNN) for contextual and temporal feature modelling, and a fully connected Softmax layer for final classification. To improve image quality and feature representation, the preprocessing pipeline incorporates bicubic interpolation, Bi-Histogram Equalization with Plateau Limit (BHEPL) for contrast enhancement, Kapur entropy-based thresholding, and the Roberts Cross operator for edge detection. The framework is evaluated on the Intel & MobileODT Cervical Cancer Screening dataset, which contains three cervix types (Type 1, Type 2, and Type 3) defined by the location and visibility of the transformation zone. CerviDBF-Net achieves an accuracy of 98.92%, sensitivity of 99.13%, specificity of 98.38%, precision of 98.69%, and an F1-score of 98.27%, demonstrating strong classification performance compared with the evaluated existing models. The proposed approach provides an automated baseline tool for cervix-type classification that may support triage in colposcopy-assisted screening by helping clinicians assess whether adequate visualization of the squamocolumnar junction has been achieved.